Additional fine tuning producing torch files

[zachNA2]

Hello, I have fine tuned the 7B model. I then attempted the following scenarios.

Scenario 1: Used the export_to_hf file to convert the new models merged together. From their I trained it on the next data set. I used the following code.

tokenizer = LlamaTokenizer.from_pretrained("decapoda-research/llama-7b-hf", add_eos_token=True)

model = LlamaForCausalLM.from_pretrained( "PATH_TO_CONVERTED_CKPT_FOLDER", load_in_8bit=False, )

I kept the rest the same (modified prompts for my data) and did the resume form checkpoint in trainer.train and pointed it to the hf_ckpt file as well as tried with the lora-alpaca folder.

Result: produced a pytorch_model.bin and training_args.bin, but nothing in the HF side I could use to inference.

Scenario 2: Tried to load model from the lora-alpaca folder.

tokenizer = LlamaTokenizer.from_pretrained("decapoda-research/llama-7b-hf", add_eos_token=True)

model = LlamaForCausalLM.from_pretrained( "PATH_TO_CONVERTED_LORA-ALPACA", load_in_8bit=False, )

Again this literally creates pytorch files...

Scenario 3:

tokenizer = LlamaTokenizer.from_pretrained("decapoda-research/llama-7b-hf", add_eos_token=True)

model = LlamaForCausalLM.from_pretrained( "decapoda-research/llama-7b-hf", load_in_8bit=False, )

Put the resume from checkpoint on ckpt folder as well as tried it with lara-alpaca.

Result: got pytorch files...

---

I've kept the same training format (same params and everything). The only thing changed is the model path... What is happening?